Corporate Bench: Comprehensive Benchmark Architecture for Corporate AI

Updated 1 September 2026

CorporateBench is a proposed multidimensional benchmark architecture designed to evaluate AI systems across a range of corporate tasks, including financial prediction,reasoning, search, and decision support.
Examples of its potential components include annual report analysis, ESG question answering, governance reasoning, and long-horizon enterprise simulation, all based on from comprehensive corpus of corporate documents and simulations.
The benchmark aims to distinguish legal settings, supporting separate tracks for financial prediction, governance reasoning, currency, asset extraction, and multi-horizon enterprise simulation, all evaluated with rigorous, context-aware metrics.

CorporateBench is a proposed comprehensive benchmark for evaluating artificial-intelligence systems on corporate information, reasoning, prediction, search, and decision-support tasks. The supplied research does not define an independently released benchmark named CorporateBench; rather, the term is used as a unifying design concept across studies of corporate bankruptcy prediction, financial-document analysis, ESG question answering, governance reasoning, enterprise simulation, people search, organizational-document extraction, and financial-services meta-evaluation. A plausible interpretation is that CorporateBench denotes a modular evaluation suite combining document-grounded reasoning, structured-data analysis, predictive modeling, agentic workflow execution, and operational-risk assessment.

1. Scope and conceptual foundations

CorporateBench is best understood as a multidimensional benchmark architecture rather than a single task or dataset. Its candidate components span several levels of corporate activity:

Foundational ability: financial concepts, numerical calculation, formula application, document extraction, and structured-data access.
Reasoning ability: comparison across periods or companies, risk assessment, financial insight generation, governance interpretation, and causal or multi-hop reasoning.
Technical skill: SQL generation, exploratory-analysis code, visualization planning, data transformation, and executable program synthesis.
Predictive modeling: bankruptcy, financial distress, corporate emissions, and other firm-level outcomes.
Operational agency: long-horizon strategy, budgeting, pricing, marketing, negotiation, information acquisition, and adaptation to changing environments.
Governance and control: charter compliance, organizational-document extraction, insider-threat detection, attribution, escalation, and evidence-based intervention.
Search and recommendation: recruiting, B2B prospecting, account discovery, expert search, and corporate-relationship qualification.

This breadth distinguishes CorporateBench from specialized benchmarks. BizBench focuses on quantitative business and finance reasoning (Koncel-Kedziorski et al., 2023); SECQUE evaluates open-ended analysis of SEC filings (Yoash et al., 6 Apr 2025); Climate Finance Bench and ESGBench focus on sustainability reports and ESG evidence grounding (Mankour et al., 28 May 2025, George et al., 20 Nov 2025); CHANCERY evaluates corporate-governance reasoning over charters (Basti et al., 6 May 2026); and CEO-Bench evaluates sustained operation of a simulated company over 500 days (Chen et al., 16 Jun 2026).

A suitable abstraction is:

$\text{Benchmark} \rightarrow \text{work activity} \rightarrow \text{enterprise function} \rightarrow \text{use case or control objective}.$

The work-activity layer can draw on O*NET Generalized Work Activities, while the enterprise-function layer can extend the BIAN banking taxonomy used in a financial-services meta-benchmark (Hudson, 2 Jul 2026). Relevant functions include Finance, Legal, Compliance, Human Resources, Sales, Marketing, Procurement, Supply Chain, Operations, IT, Security, Strategy, Corporate Relations, and Knowledge Management.

CorporateBench should not be interpreted as a certification of deployment readiness. Benchmark capability, real-world business competence, and deployment approval are distinct:

$\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$

2. Corporate documents and evidence-grounded reasoning

A central CorporateBench capability is grounded analysis of long, heterogeneous corporate documents. The relevant sources include annual reports, 10-K and 10-Q filings, sustainability reports, climate disclosures, charters, bylaws, earnings materials, governance documents, and organizational records.

Financial-document retrieval and extraction

LEDGER contains 4,999 annual reports from 738 companies covering fiscal years 2009–2024, with approximately 691,858 pages and 31 consolidated financial KPIs per company-year report (Moslonka et al., 11 Jun 2026). It derives:

page-level KPI retrieval with TREC-style relevance judgments;
full-report single-value lookup;
full KPI extraction from long OCR documents.

The benchmark preserves narrative prose, tables, figures, CEO letters, financial statements, and page-aligned OCR. Its page-retrieval baselines report Recall@5 of 0.370 for ColBERT, 0.272 for SPLADE, and 0.265 for BM25. In the long-context lookup task, Qwen3.6-27B reaches recall of 91.4 and precision of 93.5, while no evaluated model exceeds 80% recall in full 31-field KPI extraction.

These results indicate that locating a relevant page and extracting a value are separable capabilities. CorporateBench should therefore evaluate retrieval, cell or span selection, unit interpretation, fiscal-period alignment, and answer generation independently.

Open-ended filing analysis

SECQUE contains 565 expert-written questions over SEC filings, divided into comparison and trend analysis, ratio analysis, risk factors, and analyst insights (Yoash et al., 6 Apr 2025). Its questions require:

cross-period and cross-company comparison;
financial-ratio calculation;
risk interpretation;
synthesis of numerical and textual evidence;
analyst-style explanation.

The four categories contain 220 comparison questions, 188 ratio questions, 85 risk-factor questions, and 72 analyst-insight questions. The benchmark’s baseline results show that analyst insights are more difficult than risk-factor extraction, while ratio and comparison questions expose formula-selection, unit, sign, and long-context errors.

SECQUE uses source references including accession number, page number, filing section, and question type. Its three-level evaluation labels answers as correct, incomplete, or incorrect. SECQUE-Judge, based on GPT-4o, achieves class-2 F1 of 0.85, precision of 0.905, recall of 0.800, and overall accuracy of 0.75 against human judgments. The evaluator is more reliable for identifying fully acceptable answers than for distinguishing incomplete from incorrect responses.

ESG and climate disclosures

Climate Finance Bench contains 33 English-language sustainability reports from companies across all 11 GICS sectors and 330 expert-validated questions (Mankour et al., 28 May 2025). Its question categories are pure extraction, numerical reasoning, and logical reasoning. Questions concern Scope 1–3 emissions, renewable energy, carbon intensity, climate targets, transition plans, materiality, and scenario alignment.

Its retrieval experiments use dense retrieval, BM25, Reciprocal Rank Fusion, and cross-encoder reranking. The strongest reported configuration combines 75% dense retrieval and 25% BM25, followed by reranking. Correct-answer rates improve from 54.8% with minimal retrieval to 59.1% after BM25 and 62.1% after reranking. The benchmark therefore supports separate measurement of answer-bearing passage retrieval and answer quality.

ESGBench provides a smaller evidence-linked ESG QA setting with 119 QA pairs from 12 reports and 10 companies (George et al., 20 Nov 2025). Each record contains a question, answer, evidence quote, page number, company, document, category, and KPI name. Its baseline reports Exact Match of 21.0%, average token-level String F1 of 55.4%, Numeric Accuracy at $\pm 2\%$ of 45.3%, and approximate Recall@5 of 70–80%. Environmental numeric questions are particularly difficult, while strategy questions receive a reported score of 90.9%.

Evidence requirements

CorporateBench should represent evidence at several levels:

document identifier;
accession, URL, or report name;
publication or filing date;
page number;
section, table, row, and column;
exact evidence span;
evidence type;
fiscal period;
unit and scale;
scope and qualifiers;
source freshness;
evidence confidence.

A model response should be evaluated for both answer correctness and evidence validity. Useful metrics include retrieval Recall@K, MRR, nDCG, citation precision, citation recall, evidence entailment, and evidence sufficiency.

For unanswerable questions, the benchmark should distinguish:

not disclosed in the source;
not retrieved;
ambiguous;
contradicted;
unavailable because of parsing failure;
unknown or unverifiable.

The response convention “Not available in the retrieved information,” used in Climate Finance Bench, provides a useful abstention format (Mankour et al., 28 May 2025).

3. Financial prediction and structured corporate analytics

CorporateBench should contain predictive tasks in which labels, horizons, information availability, and entity membership are explicit. Bankruptcy and emissions prediction provide complementary examples.

Bankruptcy and financial distress

The bankruptcy benchmark of Alanis, Chava, and Shah contains 131,261 firm-year observations, including 2,585 bankruptcies and 128,676 non-bankruptcies, covering publicly traded U.S. firms (Alanis et al., 2022). Bankruptcy includes Chapter 11 reorganizations and Chapter 7 liquidations. The task is one-year-ahead prediction under a naturally imbalanced prevalence of approximately 1.97%.

The study evaluates LASSO and ridge logistic hazard models, classification and survival random forests, XGBoost, LightGBM, and neural networks. Its walk-forward design uses expanding training windows, two-year validation windows, and one-year test periods from 1990 through 2019. Accounting information is lagged by at least six months before the forecast year, market information is measured using the preceding year, and macroeconomic information is measured in December of the preceding year.

In the full specification, XGBoost achieves an AUC of 0.915, summarized elsewhere as approximately 0.92. Market variables—especially annual excess stock returns, idiosyncratic volatility, and relative size—are consistently important. Accounting variables add substantial predictive value, whereas industry and macroeconomic variables add little incremental AUC after firm-level information is included. Survival random forest produces the largest simulated lending profit, funding 54,302 loans and generating reported profit of 3,083.13 in thousands, although LASSO and ridge achieve higher ROA with very small market shares.

The domain-adapted BERT study adds contextual analysis of MD&A disclosures to five financial variables (Kim et al., 2023). Its dataset contains 59,389 firm-year observations, 520 bankrupt observations, and 58,867 non-bankrupt observations. Bankruptcy is defined as a first bankruptcy filing within 365 days after annual-report issuance.

The study compares dictionary sentiment, Word2Vec-CNN sentiment, FinBERT, and confidence-filtered domain-adapted FinBERT. Linear SVM with financial variables plus domain-adapted BERT achieves bankruptcy-class accuracy of 91.56% and non-bankruptcy accuracy of 93.26%. The corresponding hazard-model pseudo- $R^2$ increases from 16.23% for financial variables alone to 26.38% with domain-adapted BERT features.

These studies motivate separate CorporateBench tracks for:

legal bankruptcy;
financial distress;
default;
insolvency;
restructuring;
operational failure;
market stress.

The targets should not be conflated. Financial distress is not necessarily a legal bankruptcy filing, and a legal filing need not satisfy a composite financial-distress criterion.

Multi-horizon distress prediction

V4FinBench contains 1,106,879 company-year observations from 203,900 companies in Poland, Hungary, the Czech Republic, and Slovakia, covering 2006–2021 (Kostrzewa et al., 11 May 2026). It provides 131 financial and non-financial features and six horizons, $h\in\{0,1,2,3,4,5\}$ .

Its distress label requires simultaneous failure of solvency, profitability, and liquidity:

$\left(\frac{\text{equity}}{\text{total assets}}<0\right) \land \left(\frac{\text{EBITDA}}{\text{total assets}}<0\right) \land \left(\frac{\text{current assets}}{\text{current liabilities}}<0.6\right).$

Positive rates decline from 0.359% at the contemporaneous horizon to 0.193% at the five-year horizon. XGBoost reaches ROC-AUC of 0.996 at $h=0$ and 0.837 at $h=5$ . Prototype-undersampled TabPFN improves over unresampled TabPFN and surpasses the strongest gradient-boosting baseline on $F_1$ from $h=2$ onward in the reported results.

The benchmark uses company-level grouping, country stratification, fold-specific imputation and standardization, validation-based threshold selection, and released fold assignments. CorporateBench should adopt these controls and add strict forward-time evaluation:

$\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 0

Because positive rates are below 0.4%, ROC-AUC and $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 1 should be supplemented with PR-AUC, precision at fixed review capacity, recall at fixed false-positive rates, calibration, Brier score, and subgroup metrics.

Corporate emissions

GHGbench contains 32,830 enriched company-year rows from 12,087 companies covering 2018–2023 (Duan et al., 13 May 2026). The company track includes 31,331 usable Scope 1+2 labels, 18,763 lenient Scope 3 rows, and 9,814 strict Scope 3 rows.

Scope 1+2 is constructed as reported Scope 1 plus location-based Scope 2:

$\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 2

The benchmark predicts emissions in logarithmic space. Its feature regimes progress from country, sector, year, and ExioML factor to matched firm features including revenue, employees, market capitalization, and EBITDA, followed by SICS sub-sector target encoding.

On the strict Scope 1+2 panel, tuned XGBoost reaches $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 3, tuned random forest 0.875, tuned LightGBM 0.873, and TabPFN v2 0.857. Revenue is the most important feature, with a permutation $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 4 drop of $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 5. Cross-region performance declines substantially: Ridge reaches 0.506, random forest 0.485, LightGBM 0.478, XGBoost 0.468, and TabPFN 0.467.

The results show that in-distribution accuracy can be driven by firm-scale variables while geographic transfer remains difficult. Scope 3 is harder than Scope 1+2 because of lower coverage, heterogeneous categories, and greater measurement inconsistency. CorporateBench should therefore evaluate both fit within known reporting regimes and transfer to unseen countries, sectors, years, and company groups.

4. Governance, legal reasoning, and organizational control

CorporateBench should distinguish extraction of governance provisions from reasoning about proposed corporate actions.

Governance-document extraction

DECODEM evaluates document-level binary classification of governance variables in 150 bylaws and 300 charters randomly sampled from a comprehensive corpus of U.S. publicly traded corporations (Frankenreiter, 17 Jul 2026). Bylaws cover filings from 1995–2024, while charters cover 1995–2019.

Variables include:

director and officer indemnification;
advancement of expenses;
directors’ and officers’ insurance;
director and officer exculpation;
staggered boards;
special-meeting limitations;
written-consent limitations;
supermajority provisions;
cumulative voting;
takeover defenses;
shareholder voting arrangements.

Each document receives multiple binary labels, although each individual variable is a document-level classification problem. The benchmark compares prompt design, document handling, task decomposition, and cascading extraction pipelines. More elaborate pipelines do not consistently improve frontier-model performance, but can reduce the performance gap between frontier and efficiency-oriented models.

DECODEM is primarily an extraction benchmark. It asks whether a provision is present, not whether a proposed action is legally permissible.

Charter-grounded legal reasoning

CHANCERY contains 502 handcrafted questions over 79 corporate charters selected from a source collection of approximately 10,000 real-life charters (Basti et al., 6 May 2026). The benchmark is based on 24 governance principles, including anti-greenmail, blank-check preferred stock, classified boards, poison pills, cumulative voting, director indemnification, golden parachutes, supermajority requirements, unequal voting rights, and limitations on action by written consent.

Each item supplies a proposed action, a charter identifier, and a binary answer of “Yes” or “No.” The model must determine whether the action is consistent with applicable charter rules and, in some cases, general corporate law such as the Delaware General Corporation Law.

The reported accuracies are 55.4% for QwQ-32B, 58.6% for DeepSeek-R1, 64.5% for Claude 3.7 Sonnet, 72.7% for the reported Llama 70B model, and 75.2% for GPT-4o. ReAct reaches 76.1%, while CodeAct reaches 78.1%.

The principal failure modes are:

failure to synthesize multiple provisions;
confusion over role-specific authority;
misinterpretation of exceptions and negation;
numerical-threshold errors;
treating silence as prohibition;
failure to connect amendment procedures to the provision being amended;
overuse or underuse of external law;
confusion between account or actor identity;
over-conservative judgment on compound proposals.

CorporateBench should expand the binary format into labels such as:

permitted;
prohibited;
requires board approval;
requires shareholder approval;
governed by external law;
ambiguous;
insufficient information.

Every answer should include the relevant clause, actor, authority source, triggering condition, exception analysis, and external legal rule when applicable.

Insider-threat and organizational-control reasoning

OrgForge-IT introduces a verifiable synthetic benchmark in which a deterministic Python simulation engine maintains organizational state, event timing, behavioral flags, causal links, and ground truth, while LLMs generate only constrained surface prose (Flynt, 23 Mar 2026).

Its reported corpus contains 51 simulation days, 2,904 observable telemetry records, 106 true-positive records, 2,798 noise records, and a 96.4% noise rate. It evaluates ghost logins, vishing, three-phase host data hoarding, and trust-building social engineering across negligent, disgruntled, and malicious threat classes.

The benchmark separates:

baseline calibration;
triage;
correlation and verdict.

It reports triage F1, verdict F1, baseline false-positive rate, onset sensitivity, victim attribution, host-trail reconstruction, behavior citations, and schema adherence. Eight models achieve triage F1 of 0.80 but divide between verdict F1 of 1.00 and 0.80, showing that successful escalation does not imply correct attribution or final judgment.

This architecture is transferable to CorporateBench. A deterministic corporate state engine could generate internally consistent:

invoices;
purchase orders;
approvals;
ledger entries;
contracts;
emails;
support tickets;
HR records;
access logs;
project updates;
board materials.

LLMs could render these records into realistic prose while the engine preserves causal and temporal consistency. Ground truth could include affected entities, responsibility, temporal order, counterfactuals, and relevant evidence.

5. Long-horizon agents, search, and enterprise workflows

CorporateBench should evaluate not only static answers but also persistent decision-making and enterprise workflows.

Corporate management under uncertainty

CEO-Bench simulates a subscription-software company, NovaMind, for 500 days (Chen et al., 16 Jun 2026). The company begins with zero customers and $1,000,000 in cash. The agent operates through a programmable Python interface exposing 34 tools and a 19-table operational database.

The agent controls:

pricing and monetization;
marketing and customer acquisition;
product quality and research;
infrastructure and support;
enterprise negotiations;
information acquisition;
cash management.

The environment contains 26 customer groups, stochastic demand, latent willingness to pay, delayed revenue, competitor events, macroeconomic dynamics, satisfaction memory, reputation spillovers, and capacity constraints. Bankruptcy occurs when cash falls strictly below zero.

Claude Opus 4.8 achieves maximum final cash of $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$621,297,707 but has two bankruptcies. The rule-based baseline reaches $\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$71 million starting balance on their best runs.

The benchmark reveals the importance of:

cohort-based forecasting;
mining negotiation history;
targeted product development;
delayed-effect reasoning;
contingency planning;
information acquisition;
adaptation to competitors;
coordination between growth, quality, infrastructure, support, and cash.

Final cash should not be the sole CorporateBench objective. Additional metrics should include survival, cash-flow volatility, forecast error, customer retention, capacity failures, compliance violations, security incidents, stakeholder outcomes, strategic reversibility, and robustness across random seeds.

People and company search

PeopleSearchBench evaluates four people-search platforms on 119 real-world queries covering recruiting, B2B sales prospecting, expert search, and influencer discovery (Wang et al., 29 Mar 2026). It decomposes performance into Relevance Precision, Effective Coverage, and Information Utility.

The benchmark uses Criteria-Grounded Verification. Explicit query criteria are extracted, each returned person is checked against each criterion using external web evidence, and criterion-level judgments are assigned as met, partially met, or not met. The platform-level composite is:

$\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$8

Lessie scores 65.2 overall, Exa 55.0, Claude Code 46.0, and Juicebox 45.8. Lessie achieves 100% task completion and a mean of 10.4 qualified results per query.

CorporateBench should extend this framework from person-level search to person–company–relationship qualification. B2B results should verify:

current employment;
title and seniority;
company size and industry;
geography;
buying authority;
technology or business-initiative evidence;
account fit;
freshness of evidence;
role diversity within accounts;
qualified-account coverage.

Missing evidence should be distinguished from contradiction. A public profile lacking a verified fact is not necessarily evidence that the fact is false.

Corporate workflow execution

The workflow layer should evaluate whether an agent can transform records into actions. Candidate tasks include:

reconcile invoices, purchase orders, and payments;
identify unauthorized approvals;
detect inconsistent reporting periods;
route a contract for review;
generate SQL over an enterprise schema;
produce an exploratory-analysis notebook;
compare budgets with actuals;
identify a material disclosure risk;
map governance authority to a proposed action;
recommend escalation or abstention;
execute a multi-step financial or operational process.

Each workflow should expose intermediate artifacts, evidence, tool calls, latency, cost, and final outcomes. A correct final answer obtained through excessive false alarms or unsupported reasoning should not receive full operational credit.

6. Evaluation, reproducibility, and meta-benchmarking

CorporateBench requires a metric architecture that separates heterogeneous capabilities rather than collapsing them into one undifferentiated score.

Task-specific metrics

For classification and prediction, relevant measures include:

ROC-AUC;
PR-AUC;
precision;
recall;
$\text{Benchmark capability} \neq \text{real-world business competence} \neq \text{deployment approval}.$ 9;
specificity;
calibration;
Brier score;
recall at fixed false-positive rates;
precision at fixed review capacity;
cost-weighted utility.

For retrieval and extraction, relevant measures include:

Recall@1, @5, @10, and @20;
MRR;
nDCG;
evidence-span precision and recall;
table-cell accuracy;
numerical accuracy with unit compatibility;
fiscal-period accuracy;
citation correctness;
answer-bearing passage recall.

For open-ended analysis, evaluation should combine:

factual correctness;
completeness;
numerical correctness;
relevance;
evidence sufficiency;
citation validity;
interpretive soundness;
actionable value;
abstention quality.

For agents and workflows, metrics should include:

final cash or business outcome;
survival;
bankruptcy rate;
workload;
false-positive rate;
review volume;
latency;
tool cost;
forecast error;
intervention quality;
cross-functional consistency;
robustness across seeds and scenarios.

Exact and semantic scoring

Exact structured scoring is useful for reproducibility, but it can conflate reasoning with format compliance. OrgForge-IT separates exact-match scoring from semantic scoring, while BizBench uses executable Python for program-synthesis tasks (Flynt, 23 Mar 2026, Koncel-Kedziorski et al., 2023).

CorporateBench should provide parallel tracks:

exact track: canonical labels, schemas, SQL, numerical formats, and required fields;
semantic track: meaning-preserving paraphrases, equivalent formulas, valid query variants, and evidence-grounded explanations.

For numerical answers, the benchmark should store:

original surface form;
normalized value;
unit;
scale;
period;
scope;
denominator;
qualifiers;
acceptable alternative calculations.

Dynamic aggregation

The meta-benchmarking framework for financial services maps 452 benchmark identifiers to 41 O*NET work activities and 38 BIAN business domains (Hudson, 2 Jul 2026). It weights benchmarks by discrimination, coverage, and recency:

$\pm 2\%$ 0

It then uses weighted pairwise Elo updates:

$\pm 2\%$ 1

This approach avoids direct normalization of incompatible raw scores and downweights saturated, stale, or sparsely reported benchmarks. CorporateBench could adopt the architecture while adding:

enterprise-function mappings beyond banking;
risk-critical capabilities;
modality tags;
jurisdiction;
confidentiality;
data sensitivity;
tool dependence;
consequence of failure;
evidence strength;
benchmark correlation clusters.

The framework should report raw task scores, work-activity profiles, enterprise-function scores, benchmark weights, evidence density, missingness, and uncertainty separately. A composite score should never obscure a critical failure in factuality, privacy, compliance, attribution, or authorization.

Data splits and leakage control

CorporateBench should use different split strategies for different tasks:

Company-level splits to prevent repeated-company leakage.
Document-level splits to prevent near-duplicate report contamination.
Forward-time splits for deployment-oriented prediction.
Country-held-out splits for institutional transfer.
Sector-held-out splits for industry generalization.
Charter-level splits for governance reasoning.
Organization-level splits for synthetic enterprise simulations.
Private or rotating test sets for contamination resistance.

All imputers, scalers, PCA transformations, target encodings, retrieval indexes, feature-selection procedures, and model-tuning operations should be fit only on permitted training data. Filing text should be aligned by actual publication date rather than fiscal year alone. Domain adaptation should distinguish inductive training-period-only adaptation from transductive use of unlabeled test-period documents.

Reproducibility requirements

A CorporateBench release should include:

document manifests;
source URLs or accessions;
checksums and data vintages;
identifier crosswalks;
company and document metadata;
event and label definitions;
feature formulas;
unit and currency rules;
missing-value and outlier policies;
train, validation, and test manifests;
exact prompts;
model and software versions;
random seeds;
hyperparameter grids;
selected parameters;
retrieval configurations;
parsing and normalization code;
prediction files;
evaluation scripts;
confidence intervals;
economic-simulation assumptions;
deviations from reference studies.

Environmental reporting should include hardware, runtime, token counts, model precision, API or local execution, indexing cost, evaluation cost, carbon-intensity assumptions, and uncertainty. Climate Finance Bench reports API and local emissions estimates and highlights the effect of 4-bit quantization on memory and estimated emissions (Mankour et al., 28 May 2025).

CorporateBench should also document licensing and privacy constraints. Public corporate documents can contain copyrighted material, confidential information, personal data, or restrictions on redistribution. Search and recruiting components require additional treatment of privacy, source freshness, fairness, and sensitive-attribute inference.

CorporateBench is therefore most coherently defined as a modular, evidence-grounded evaluation ecosystem for corporate AI. Its foundation should combine document retrieval and extraction, financial prediction, ESG and climate analysis, governance reasoning, organizational control, people and company search, executable data analysis, and long-horizon enterprise simulation. The central methodological principles are separation of retrieval from reasoning, separation of extraction from judgment, explicit temporal and entity controls, deterministic verification where possible, calibrated abstention, operationally meaningful metrics, and transparent evidence provenance.